1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * processor_perflib.c - ACPI Processor P-States Library ($Revision: 71 $)
5 * Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
6 * Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
7 * Copyright (C) 2004 Dominik Brodowski <linux@brodo.de>
8 * Copyright (C) 2004 Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
9 * - Added processor hotplug support
12 #include <linux/kernel.h>
13 #include <linux/module.h>
14 #include <linux/init.h>
15 #include <linux/cpufreq.h>
16 #include <linux/slab.h>
17 #include <linux/acpi.h>
18 #include <acpi/processor.h>
20 #include <asm/cpufeature.h>
23 #define PREFIX "ACPI: "
25 #define ACPI_PROCESSOR_CLASS "processor"
26 #define ACPI_PROCESSOR_FILE_PERFORMANCE "performance"
27 #define _COMPONENT ACPI_PROCESSOR_COMPONENT
28 ACPI_MODULE_NAME("processor_perflib");
30 static DEFINE_MUTEX(performance_mutex);
33 * _PPC support is implemented as a CPUfreq policy notifier:
34 * This means each time a CPUfreq driver registered also with
35 * the ACPI core is asked to change the speed policy, the maximum
36 * value is adjusted so that it is within the platform limit.
38 * Also, when a new platform limit value is detected, the CPUfreq
39 * policy is adjusted accordingly.
43 * -1 -> cpufreq low level drivers not initialized -> _PSS, etc. not called yet
45 * 0 -> cpufreq low level drivers initialized -> consider _PPC values
46 * 1 -> ignore _PPC totally -> forced by user through boot param
48 static int ignore_ppc = -1;
49 module_param(ignore_ppc, int, 0644);
50 MODULE_PARM_DESC(ignore_ppc, "If the frequency of your machine gets wrongly" \
51 "limited by BIOS, this should help");
53 static bool acpi_processor_ppc_in_use;
55 static int acpi_processor_get_platform_limit(struct acpi_processor *pr)
57 acpi_status status = 0;
58 unsigned long long ppc = 0;
65 * _PPC indicates the maximum state currently supported by the platform
66 * (e.g. 0 = states 0..n; 1 = states 1..n; etc.
68 status = acpi_evaluate_integer(pr->handle, "_PPC", NULL, &ppc);
70 if (status != AE_NOT_FOUND)
71 acpi_processor_ppc_in_use = true;
73 if (ACPI_FAILURE(status) && status != AE_NOT_FOUND) {
74 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PPC"));
78 pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr->id,
79 (int)ppc, ppc ? "" : "not");
81 pr->performance_platform_limit = (int)ppc;
83 if (ppc >= pr->performance->state_count ||
84 unlikely(!dev_pm_qos_request_active(&pr->perflib_req)))
87 ret = dev_pm_qos_update_request(&pr->perflib_req,
88 pr->performance->states[ppc].core_frequency * 1000);
90 pr_warn("Failed to update perflib freq constraint: CPU%d (%d)\n",
97 #define ACPI_PROCESSOR_NOTIFY_PERFORMANCE 0x80
99 * acpi_processor_ppc_ost: Notify firmware the _PPC evaluation status
100 * @handle: ACPI processor handle
101 * @status: the status code of _PPC evaluation
102 * 0: success. OSPM is now using the performance state specificed.
103 * 1: failure. OSPM has not changed the number of P-states in use
105 static void acpi_processor_ppc_ost(acpi_handle handle, int status)
107 if (acpi_has_method(handle, "_OST"))
108 acpi_evaluate_ost(handle, ACPI_PROCESSOR_NOTIFY_PERFORMANCE,
112 void acpi_processor_ppc_has_changed(struct acpi_processor *pr, int event_flag)
116 if (ignore_ppc || !pr->performance) {
118 * Only when it is notification event, the _OST object
119 * will be evaluated. Otherwise it is skipped.
122 acpi_processor_ppc_ost(pr->handle, 1);
126 ret = acpi_processor_get_platform_limit(pr);
128 * Only when it is notification event, the _OST object
129 * will be evaluated. Otherwise it is skipped.
133 acpi_processor_ppc_ost(pr->handle, 1);
135 acpi_processor_ppc_ost(pr->handle, 0);
138 cpufreq_update_limits(pr->id);
141 int acpi_processor_get_bios_limit(int cpu, unsigned int *limit)
143 struct acpi_processor *pr;
145 pr = per_cpu(processors, cpu);
146 if (!pr || !pr->performance || !pr->performance->state_count)
148 *limit = pr->performance->states[pr->performance_platform_limit].
149 core_frequency * 1000;
152 EXPORT_SYMBOL(acpi_processor_get_bios_limit);
154 void acpi_processor_ignore_ppc_init(void)
160 void acpi_processor_ppc_init(int cpu)
162 struct acpi_processor *pr = per_cpu(processors, cpu);
165 ret = dev_pm_qos_add_request(get_cpu_device(cpu),
166 &pr->perflib_req, DEV_PM_QOS_MAX_FREQUENCY,
169 pr_err("Failed to add freq constraint for CPU%d (%d)\n", cpu,
175 void acpi_processor_ppc_exit(int cpu)
177 struct acpi_processor *pr = per_cpu(processors, cpu);
179 dev_pm_qos_remove_request(&pr->perflib_req);
182 static int acpi_processor_get_performance_control(struct acpi_processor *pr)
185 acpi_status status = 0;
186 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
187 union acpi_object *pct = NULL;
188 union acpi_object obj = { 0 };
191 status = acpi_evaluate_object(pr->handle, "_PCT", NULL, &buffer);
192 if (ACPI_FAILURE(status)) {
193 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PCT"));
197 pct = (union acpi_object *)buffer.pointer;
198 if (!pct || (pct->type != ACPI_TYPE_PACKAGE)
199 || (pct->package.count != 2)) {
200 printk(KERN_ERR PREFIX "Invalid _PCT data\n");
209 obj = pct->package.elements[0];
211 if ((obj.type != ACPI_TYPE_BUFFER)
212 || (obj.buffer.length < sizeof(struct acpi_pct_register))
213 || (obj.buffer.pointer == NULL)) {
214 printk(KERN_ERR PREFIX "Invalid _PCT data (control_register)\n");
218 memcpy(&pr->performance->control_register, obj.buffer.pointer,
219 sizeof(struct acpi_pct_register));
225 obj = pct->package.elements[1];
227 if ((obj.type != ACPI_TYPE_BUFFER)
228 || (obj.buffer.length < sizeof(struct acpi_pct_register))
229 || (obj.buffer.pointer == NULL)) {
230 printk(KERN_ERR PREFIX "Invalid _PCT data (status_register)\n");
235 memcpy(&pr->performance->status_register, obj.buffer.pointer,
236 sizeof(struct acpi_pct_register));
239 kfree(buffer.pointer);
246 * Some AMDs have 50MHz frequency multiples, but only provide 100MHz rounding
247 * in their ACPI data. Calculate the real values and fix up the _PSS data.
249 static void amd_fixup_frequency(struct acpi_processor_px *px, int i)
251 u32 hi, lo, fid, did;
252 int index = px->control & 0x00000007;
254 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
257 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
258 || boot_cpu_data.x86 == 0x11) {
259 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
262 * Bit 63: PstateEn. Read-write. If set, the P-state is valid.
269 if (boot_cpu_data.x86 == 0x10)
270 px->core_frequency = (100 * (fid + 0x10)) >> did;
272 px->core_frequency = (100 * (fid + 8)) >> did;
276 static void amd_fixup_frequency(struct acpi_processor_px *px, int i) {};
279 static int acpi_processor_get_performance_states(struct acpi_processor *pr)
282 acpi_status status = AE_OK;
283 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
284 struct acpi_buffer format = { sizeof("NNNNNN"), "NNNNNN" };
285 struct acpi_buffer state = { 0, NULL };
286 union acpi_object *pss = NULL;
288 int last_invalid = -1;
291 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
292 if (ACPI_FAILURE(status)) {
293 ACPI_EXCEPTION((AE_INFO, status, "Evaluating _PSS"));
297 pss = buffer.pointer;
298 if (!pss || (pss->type != ACPI_TYPE_PACKAGE)) {
299 printk(KERN_ERR PREFIX "Invalid _PSS data\n");
304 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d performance states\n",
305 pss->package.count));
307 pr->performance->state_count = pss->package.count;
308 pr->performance->states =
309 kmalloc_array(pss->package.count,
310 sizeof(struct acpi_processor_px),
312 if (!pr->performance->states) {
317 for (i = 0; i < pr->performance->state_count; i++) {
319 struct acpi_processor_px *px = &(pr->performance->states[i]);
321 state.length = sizeof(struct acpi_processor_px);
324 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Extracting state %d\n", i));
326 status = acpi_extract_package(&(pss->package.elements[i]),
328 if (ACPI_FAILURE(status)) {
329 ACPI_EXCEPTION((AE_INFO, status, "Invalid _PSS data"));
331 kfree(pr->performance->states);
335 amd_fixup_frequency(px, i);
337 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
338 "State [%d]: core_frequency[%d] power[%d] transition_latency[%d] bus_master_latency[%d] control[0x%x] status[0x%x]\n",
340 (u32) px->core_frequency,
342 (u32) px->transition_latency,
343 (u32) px->bus_master_latency,
344 (u32) px->control, (u32) px->status));
347 * Check that ACPI's u64 MHz will be valid as u32 KHz in cpufreq
349 if (!px->core_frequency ||
350 ((u32)(px->core_frequency * 1000) !=
351 (px->core_frequency * 1000))) {
352 printk(KERN_ERR FW_BUG PREFIX
353 "Invalid BIOS _PSS frequency found for processor %d: 0x%llx MHz\n",
354 pr->id, px->core_frequency);
355 if (last_invalid == -1)
358 if (last_invalid != -1) {
360 * Copy this valid entry over last_invalid entry
362 memcpy(&(pr->performance->states[last_invalid]),
363 px, sizeof(struct acpi_processor_px));
369 if (last_invalid == 0) {
370 printk(KERN_ERR FW_BUG PREFIX
371 "No valid BIOS _PSS frequency found for processor %d\n", pr->id);
373 kfree(pr->performance->states);
374 pr->performance->states = NULL;
377 if (last_invalid > 0)
378 pr->performance->state_count = last_invalid;
381 kfree(buffer.pointer);
386 int acpi_processor_get_performance_info(struct acpi_processor *pr)
390 if (!pr || !pr->performance || !pr->handle)
393 if (!acpi_has_method(pr->handle, "_PCT")) {
394 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
395 "ACPI-based processor performance control unavailable\n"));
399 result = acpi_processor_get_performance_control(pr);
403 result = acpi_processor_get_performance_states(pr);
407 /* We need to call _PPC once when cpufreq starts */
409 result = acpi_processor_get_platform_limit(pr);
414 * Having _PPC but missing frequencies (_PSS, _PCT) is a very good hint that
415 * the BIOS is older than the CPU and does not know its frequencies
419 if (acpi_has_method(pr->handle, "_PPC")) {
420 if(boot_cpu_has(X86_FEATURE_EST))
421 printk(KERN_WARNING FW_BUG "BIOS needs update for CPU "
422 "frequency support\n");
427 EXPORT_SYMBOL_GPL(acpi_processor_get_performance_info);
429 int acpi_processor_pstate_control(void)
433 if (!acpi_gbl_FADT.smi_command || !acpi_gbl_FADT.pstate_control)
436 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
437 "Writing pstate_control [0x%x] to smi_command [0x%x]\n",
438 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
440 status = acpi_os_write_port(acpi_gbl_FADT.smi_command,
441 (u32)acpi_gbl_FADT.pstate_control, 8);
442 if (ACPI_SUCCESS(status))
445 ACPI_EXCEPTION((AE_INFO, status,
446 "Failed to write pstate_control [0x%x] to smi_command [0x%x]",
447 acpi_gbl_FADT.pstate_control, acpi_gbl_FADT.smi_command));
451 int acpi_processor_notify_smm(struct module *calling_module)
453 static int is_done = 0;
456 if (!acpi_processor_cpufreq_init)
459 if (!try_module_get(calling_module))
462 /* is_done is set to negative if an error occurred,
463 * and to postitive if _no_ error occurred, but SMM
464 * was already notified. This avoids double notification
465 * which might lead to unexpected results...
468 module_put(calling_module);
470 } else if (is_done < 0) {
471 module_put(calling_module);
477 result = acpi_processor_pstate_control();
479 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No SMI port or pstate_control\n"));
480 module_put(calling_module);
484 module_put(calling_module);
488 /* Success. If there's no _PPC, we need to fear nothing, so
489 * we can allow the cpufreq driver to be rmmod'ed. */
492 if (!acpi_processor_ppc_in_use)
493 module_put(calling_module);
498 EXPORT_SYMBOL(acpi_processor_notify_smm);
500 int acpi_processor_get_psd(acpi_handle handle, struct acpi_psd_package *pdomain)
503 acpi_status status = AE_OK;
504 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
505 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"};
506 struct acpi_buffer state = {0, NULL};
507 union acpi_object *psd = NULL;
509 status = acpi_evaluate_object(handle, "_PSD", NULL, &buffer);
510 if (ACPI_FAILURE(status)) {
514 psd = buffer.pointer;
515 if (!psd || (psd->type != ACPI_TYPE_PACKAGE)) {
516 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
521 if (psd->package.count != 1) {
522 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
527 state.length = sizeof(struct acpi_psd_package);
528 state.pointer = pdomain;
530 status = acpi_extract_package(&(psd->package.elements[0]),
532 if (ACPI_FAILURE(status)) {
533 printk(KERN_ERR PREFIX "Invalid _PSD data\n");
538 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) {
539 printk(KERN_ERR PREFIX "Unknown _PSD:num_entries\n");
544 if (pdomain->revision != ACPI_PSD_REV0_REVISION) {
545 printk(KERN_ERR PREFIX "Unknown _PSD:revision\n");
550 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL &&
551 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY &&
552 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) {
553 printk(KERN_ERR PREFIX "Invalid _PSD:coord_type\n");
558 kfree(buffer.pointer);
561 EXPORT_SYMBOL(acpi_processor_get_psd);
563 int acpi_processor_preregister_performance(
564 struct acpi_processor_performance __percpu *performance)
569 cpumask_var_t covered_cpus;
570 struct acpi_processor *pr;
571 struct acpi_psd_package *pdomain;
572 struct acpi_processor *match_pr;
573 struct acpi_psd_package *match_pdomain;
575 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL))
578 mutex_lock(&performance_mutex);
581 * Check if another driver has already registered, and abort before
582 * changing pr->performance if it has. Check input data as well.
584 for_each_possible_cpu(i) {
585 pr = per_cpu(processors, i);
587 /* Look only at processors in ACPI namespace */
591 if (pr->performance) {
596 if (!performance || !per_cpu_ptr(performance, i)) {
602 /* Call _PSD for all CPUs */
603 for_each_possible_cpu(i) {
604 pr = per_cpu(processors, i);
608 pr->performance = per_cpu_ptr(performance, i);
609 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
610 pdomain = &(pr->performance->domain_info);
611 if (acpi_processor_get_psd(pr->handle, pdomain)) {
620 * Now that we have _PSD data from all CPUs, lets setup P-state
623 for_each_possible_cpu(i) {
624 pr = per_cpu(processors, i);
628 if (cpumask_test_cpu(i, covered_cpus))
631 pdomain = &(pr->performance->domain_info);
632 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
633 cpumask_set_cpu(i, covered_cpus);
634 if (pdomain->num_processors <= 1)
637 /* Validate the Domain info */
638 count_target = pdomain->num_processors;
639 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
640 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
641 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
642 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_HW;
643 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
644 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ANY;
646 for_each_possible_cpu(j) {
650 match_pr = per_cpu(processors, j);
654 match_pdomain = &(match_pr->performance->domain_info);
655 if (match_pdomain->domain != pdomain->domain)
658 /* Here i and j are in the same domain */
660 if (match_pdomain->num_processors != count_target) {
665 if (pdomain->coord_type != match_pdomain->coord_type) {
670 cpumask_set_cpu(j, covered_cpus);
671 cpumask_set_cpu(j, pr->performance->shared_cpu_map);
674 for_each_possible_cpu(j) {
678 match_pr = per_cpu(processors, j);
682 match_pdomain = &(match_pr->performance->domain_info);
683 if (match_pdomain->domain != pdomain->domain)
686 match_pr->performance->shared_type =
687 pr->performance->shared_type;
688 cpumask_copy(match_pr->performance->shared_cpu_map,
689 pr->performance->shared_cpu_map);
694 for_each_possible_cpu(i) {
695 pr = per_cpu(processors, i);
696 if (!pr || !pr->performance)
699 /* Assume no coordination on any error parsing domain info */
701 cpumask_clear(pr->performance->shared_cpu_map);
702 cpumask_set_cpu(i, pr->performance->shared_cpu_map);
703 pr->performance->shared_type = CPUFREQ_SHARED_TYPE_ALL;
705 pr->performance = NULL; /* Will be set for real in register */
709 mutex_unlock(&performance_mutex);
710 free_cpumask_var(covered_cpus);
713 EXPORT_SYMBOL(acpi_processor_preregister_performance);
716 acpi_processor_register_performance(struct acpi_processor_performance
717 *performance, unsigned int cpu)
719 struct acpi_processor *pr;
721 if (!acpi_processor_cpufreq_init)
724 mutex_lock(&performance_mutex);
726 pr = per_cpu(processors, cpu);
728 mutex_unlock(&performance_mutex);
732 if (pr->performance) {
733 mutex_unlock(&performance_mutex);
737 WARN_ON(!performance);
739 pr->performance = performance;
741 if (acpi_processor_get_performance_info(pr)) {
742 pr->performance = NULL;
743 mutex_unlock(&performance_mutex);
747 mutex_unlock(&performance_mutex);
751 EXPORT_SYMBOL(acpi_processor_register_performance);
753 void acpi_processor_unregister_performance(unsigned int cpu)
755 struct acpi_processor *pr;
757 mutex_lock(&performance_mutex);
759 pr = per_cpu(processors, cpu);
761 mutex_unlock(&performance_mutex);
766 kfree(pr->performance->states);
767 pr->performance = NULL;
769 mutex_unlock(&performance_mutex);
774 EXPORT_SYMBOL(acpi_processor_unregister_performance);